Big Data, Little Cluster: Using a Small Footprint of GPU Servers to - - PowerPoint PPT Presentation

Big Data, Little Cluster: Using a Small Footprint of GPU Servers to Interactively Query and Visualize Massive Datasets

May 9, 2017 Todd Mostak | Co-founder + CEO, MapD @toddmostak | @mapd

The data explosion is just beginning

2

Enterprise Data VOIP Social Media & Web Sensor + Devices

0k 10k 20k 30k 40k

2014 2015 2016 2017 2018 2019 2020

Source: IDC and EMC Digital Universe Report 20k 30k 10k 40k

Doubling in less than 3 years Exabytes

But storage is not the problem

3

0.00 0.02 0.04 0.06 0.08 0.10 0.12

2015 2016 2017 2018 2019 2020

Source: Wikibon 2015 4-year costs/TB SSD including packaging, power, cooling, maintenance + space

Terabytes

Amount of Storage $1 Buys (in T erabytes)

A compute inflection point

4

Data Growth CPU Processing Power

40% per year 20% per year

GPUs offer a way forward

5

Data Growth CPU Processing Power GPU Processing Power

50% per year 40% per year 20% per year

Ability to Read Data

1,000 2,000 3,000 4,000 5,000 6,000 7,000

2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

Memory Bandwidth

10 20 30 40 50 60 70 80 90

2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

Teraflops

memory bandwidth GB/sec floating point operations /sec

GPUs outperform CPUs in data critical areas

6

GPU GPU CPU CPU Compute Power Compute Power Ability to Read Data

7

MapD Core MapD Immerse

MapD: software optimized for the fastest hardware

An in-memory, relational, column store database powered by GPUs A visual analytics engine that leverages the speed + rendering capabilities of MapD Core

+

100x Faster Queries Speed of Thought Visualization

Tableau or 3rd party viz Non-Viz Output

JDBC/Hadoop

8

Where MapD sits

Streaming Data

MapD Immerse MapD Core

GPU Acceleration

JDBC, ODBC, Thrift

Kafka

MapD Core

9

The world's fastest in-memory GPU database powers the world's most immersive data exploration experience

Data Lake/Data Warehouse/SOR

Performance starts with memory management

10

SSD or NVRAM STORAGE (L3) 250GB to 20TB 1-2 GB/sec CPU RAM (L2) 32GB to 3TB 70-120 GB/sec GPU RAM (L1) 24GB to 384GB 3000-5000 GB/sec Hot Data Speedup = 1500x to 5000x Over Cold Data Warm Data Speedup = 35x to 120x Over Cold Data Cold Data COMPUTE LAYER STORAGE LAYER

SPEED INCREASES

Query Compilation with LLVM

Traditional DBs can be highly inefficient

• each operator in SQL treated as a separate function
• incurs tremendous overhead and prevents vectorization

MapD compiles queries w/LLVM to create one custom function

• Queries run at speeds approaching hand-written functions
• LLVM enables generic targeting of different architectures (GPUs, X86, ARM, etc).
• Code can be generated to run query on CPU and GPU simultaneously

10111010101001010110101101010101 00110101101101010101010101011101 11

These innovations drive exceptional speed + scale

12

Noted DB blogger, Mark Litwintschik has benchmarked MapD vs. major CPU systems and found it to be between 74x to 3,500x faster than CPU DBs.

The GPU Open Analytics Initiative (GOAI) and the GPU Data Frame (GDF)

13

ML frameworks GPU Acceleration Zone Custom functions Result set

End-to-end on the GPU: Supporting ML with MapD Compute Engine (Roadmap)

14

Output result set

MapD Immerse

Lightning fast visual analytics for the MapD Core database

Basic charts are frontend rendered using D3 and other related toolkits Scatterplots, pointmaps + polygons are backend rendered using the Iris Rendering Engine on GPUs Geo-Viz is composited over a frontend rendered basemap

MapD Immerse: our hybrid approach

16

Server side rendering

Data goes from compute (CUDA) to graphics (OpenGL) pipeline without copy and comes back as compressed PNG (~100 KB) rather than raw data (> 1GB) Vega Spec (a visualization grammar)

• A declarative JSON format for creating

visualization designs

• Used to describe backend visualizations
• Defines attributes of render primitives

which can be driven by data columns and mapped by scales

Shader Compilation Framework

• Templatized: supports multiple types (ints,

floats, colors, etc), and multiple continuities (discrete, continuous)

Backend

Query-to- Render

PNG SQL+ Vega Frontend

The X-Factor

17

Scale Up then Out

Performantly scaling the MapD Analytics Platform to analyze big data on small clusters

Benefits of a Distributed System

• Better Ability to Scale
• Multiple servers means ability to support more GPU RAM and CPU

RAM for caching bigger datasets in memory

• Better Write Performance
• The MapD 3.0 distributed capability supports distributed loading for

better throughput

• Better Read Performance
• Multiple servers can support more GPUs

19

20

“MapD was already the fastest analytics database I had ever tested, even when comparing a single server of MapD against large clusters of CPU-based solutions. With the new distributed architecture, MapD offers users record-beating performance

• ver even more massive data sets.”

Mark Litwintschik, tech.marksblogg.com

Distributed MapD Core Database Architecture

Confidential & Proprietary 21

MapD Aggregator Cluster Metadata MapD Leaf Data Data Data MapD Leaf Data Data Data MapD Leaf Data Data Data

GPUN GPU2

22 Identify and Load Data (If Needed)

Execute Query

Identify and Load Data (If Needed)

Execute Query GPU1 Accept Query Prepare Execution

Identify and Load Data (If Needed)

Execute Query Reduce Result

MapD Handler

...

Query Done?

MapD Core Database (Single Node)

Return Result

Yes No

Leaf 2 Leaf 1

Distributed Scale Up and Scale Out

. . .

Parse and Validate SQL Generate Algebraic Sequence

23 Reduce Result Query Done? Return Result

Yes

No

Prepare Execution

Leaf N Shared Dictionary

Leaf Aggregator

Distributed Benchmark

1.1B record NYC Taxi Dataset benchmark (conducted by Mark Litwintschik)

24

Query AWS P2.8xlarge timings (seconds) 2 x P2.8xlarge cluster timings (seconds) SELECT cab_type, count(*) FROM trips GROUP BY cab_type;

0.022 0.034

SELECT passenger_count, avg(total_amount) FROM trips GROUP BY passenger_count;

0.156 0.061

SELECT passenger_count, extract(year from pickup_datetime) AS pickup_year, count(*) FROM trips GROUP BY passenger_count, pickup_year;

0.309 0.178

SELECT passenger_count, extract(year from pickup_datetime) AS pickup_year, cast(trip_distance as int) AS distance, count(*) AS the_count FROM trips GROUP BY passenger_count, pickup_year, distance ORDER BY pickup_year, the_count desc;

0.771 0.499

Load Time

48 minutes 26 minutes

Demo

25

26

MapD, Now Open Source

Delivering significant customer value across industries

27

Polling smartphones on demand to assess network health. Running complex queries in real-time for customers to drive insights and ad-buys. npm looks at over 8B records at a given moment to identify trends, segments + anomalies in the javascript world.

Previously had to respond in 24+ hours. Took hours on Oracle previously. Splunk couldn’t scale economically

• r performance-wise.

Closing thoughts

28

We are at an inflection point in compute and GPUs are set to dominate the coming decade.

Closing thoughts

29

GPUs allow users to scale up before needing to scale out. lowering performance-killing network overheads and decreasing hardware and administration costs.

Closing thoughts

30

Integrated Analytics on GPUs comprising querying, viz and ML provide critical efficiencies and capabilities not found in siloed systems.